Rationalized deep learning super-resolution microscopy for sustained live imaging of rapid subcellular processes.

Area of Science:

• Optical microscopy
• Biophysics
• Computational imaging

Background:

• Deep neural networks (DNNs) advance optical microscopy for super-resolution and image restoration.
• DNNs in microscopy can introduce artifacts, limiting their application.
• Minimizing invasiveness while maximizing spatiotemporal information is key for bioprocess imaging.

Purpose of the Study:

• To develop rationalized deep learning (rDL) for structured illumination microscopy (SIM) and lattice light sheet microscopy (LLSM).
• To incorporate prior knowledge of illumination patterns into deep learning models for denoising raw microscopy images.
• To improve image quality, reduce artifacts, and enable self-supervised training in advanced microscopy techniques.

Main Methods:

• Developed rationalized deep learning (rDL) by integrating illumination pattern knowledge into DNNs.
• Applied rDL to structured illumination microscopy (SIM) to denoise raw images and correct spectral bias.
• Implemented rDL for lattice light sheet microscopy (LLSM) enabling self-supervised training using data continuity.

Main Results:

• rDL-SIM eliminated spectral bias-induced resolution degradation, reducing model uncertainty five-fold.
• Super-resolution information in rDL-SIM improved more than ten-fold compared to other computational methods.
• rDL-LLSM achieved results comparable to supervised methods using self-supervised training.

Conclusions:

• Rationalized deep learning (rDL) significantly enhances image quality and information content in optical microscopy.
• rDL provides a robust framework for artifact reduction and improved super-resolution in SIM and LLSM.
• The developed rDL approach enables advanced imaging of dynamic biological processes with high fidelity.